Text to Speech with Emotion: A 2026 Engineer's Guide to Expressive TTS

TL;DR

• Text to speech with emotion in 2026 is no longer a checkbox — it is the default. GPT-Realtime-2 ships with dynamic prosody, named emotions, and per-phrase tone steering.
• The big shift: TTS is now generative, not concatenative. You steer with natural language ("warm, calming, slow") instead of phoneme tags.
• CallSphere uses emotion-aware TTS across all 6 voice agents — the healthcare agent is warm and slow, the sales agent is upbeat and confident, the after-hours agent is calm and reassuring.
• $149-$1,499/mo, 14-day free trial, 3-5 day setup.

This is part of our Siri Voice Generator guide.

What does text to speech with emotion actually mean in 2026?

Text to speech with emotion in 2026 means a TTS system that can speak the same words with materially different prosody, pacing, pitch, and tonal warmth based on a directive — either a natural-language instruction or a structured emotion tag. The output sounds like a person who is actually feeling something, not a flat voice with pitch contour bolted on.

This is a real shift. As recently as 2023, "emotional TTS" was mostly SSML hacks — phoneme-level pitch adjustments and rate changes layered on a concatenative engine. The result sounded uncanny. In 2026, the leading models — OpenAI's GPT-Realtime-2 stack, ElevenLabs v3, Cartesia Sonic-2 — generate audio end-to-end and accept emotion directives in plain English. You write "speak this warmly and slowly, as if reassuring a worried patient" and the model does it.

I use this every day in CallSphere. The healthcare agent uses a warm, slow voice tuned for elderly callers. The sales outbound agent uses an upbeat, confident voice. The after-hours emergency escalation agent uses a calm, reassuring voice that drops the caller's heart rate, not raises it. None of those voices use the same prosody preset.

How does GPT-Realtime-2 handle text to speech with emotion?

GPT-Realtime-2, which OpenAI shipped on May 7, 2026, embeds emotion steering directly into the system prompt and the per-turn instruction. The model accepts directives like "speak with quiet enthusiasm" or "convey gentle empathy" and adjusts prosody, pacing, and pitch contour automatically. There is no separate emotion tag layer.

The pricing matters for production teams: audio input at $32 per 1M tokens, audio output at $64 per 1M tokens, cached input at $0.40 per 1M tokens. If your system prompt includes a 600-token emotion specification, cached input makes it negligible across thousands of calls.

The 128K context window means you can include long emotion playbooks — "if the caller raises their voice, switch to a calm, low register; if the caller jokes, match the warmth" — without trimming. We do exactly this in CallSphere's prompts.

What is the best speech to text model for emotional voice agents?

The speech to text side matters as much as the TTS side. The agent needs to hear the caller's emotion to respond appropriately. The leading 2026 models for this are GPT-Realtime-Whisper (OpenAI), Deepgram Nova-3, and AssemblyAI Universal-2. All three transcribe with paralinguistic features — pace, volume, hesitation markers — which the agent's logic layer uses to choose the next emotion directive.

CallSphere uses GPT-Realtime-Whisper for STT and GPT-Realtime-2 for TTS. The pipeline runs in roughly 600ms turn latency on a 4G connection, which is the floor for natural conversation. Slower than 800ms and callers start to interrupt.

For specific voice profiles — woman text to speech, australian text to speech, adam text to speech, text to speech with characters — the new generative models render any voice profile from a sample or a description. CallSphere ships with 60+ voice profiles across 57+ languages, and you can clone a custom voice from a 30-second sample if your contract allows it.

How do I steer text to speech for specific accents and personas?

You write a natural-language directive. Examples we use in production:

• Healthcare agent (default) — "Speak warmly and at a measured pace. Pause briefly between concepts. Voice should feel like a reassuring nurse, not a salesperson."
• Real estate agent — "Speak with confident energy. Pace should be upbeat but clear. Voice should feel like a top-producing agent who knows the market."
• Sales outbound agent — "Speak with quiet enthusiasm. Avoid hard-sell tone. Voice should feel like a peer recommending something useful."
• After-hours emergency agent — "Speak calmly and slowly. Lower the pitch slightly. Voice should reduce the caller's stress, not amplify it."
• Salon booking agent — "Speak with warm professionalism. Voice should feel like a friendly receptionist who remembers you."
• Hotel concierge agent — "Speak with polished hospitality. Voice should feel like a five-star concierge."

For accent-specific use cases — australian text to speech, japanese text to speech (text to speech japanese), british english — you specify the accent and locale in the voice config. The model renders it natively. You do not need a separate accent layer.

How CallSphere does this in production

Our TTS stack runs on GPT-Realtime-2 with per-agent prompts:

• 6 live agents — each has a tuned voice persona with emotion directives baked into the system prompt.
• 60+ voice profiles — male, female, multiple accents per language, multiple personas per accent.
• 57+ languages — emotion directives work across all of them; the model adapts prosody per language.
• Sub-800ms turn latency — STT + reasoning + TTS round trip on a 4G connection.
• Per-turn emotion steering — the agent can shift emotion mid-call based on caller cues (e.g., shift to a calmer voice if the caller's pitch rises).
• Custom voice cloning — 30-second sample, BYOC support, contractual licensing required.

Hear all 6 CallSphere voices in the live demo →

A real example walk-through

A dental group in Boston deployed CallSphere's healthcare agent in March 2026 for inbound appointment booking and after-hours triage. The first week, their highest-volume call type was elderly patients calling about appointment confirmations.

We tuned the agent's voice to a warm, slow, lower-register profile based on patient feedback. Average call duration on confirmation calls dropped from 3:40 to 2:10 because the agent's pacing matched the caller's. CSAT on those calls went up 18 points. The clinic's office manager called it "the first AI that sounds like it actually cares."

The mechanism is just prompt-level emotion steering: "Speak warmly, slowly, and with the patience of a 30-year veteran nurse." The model rendered that into prosody. No SSML, no custom voice training, no $50K voice budget.

Pricing & how to try it

Text to speech with emotion is included in every CallSphere plan:

• Starter — $149/mo — 2,000 interactions, all 60+ voices, all 57+ languages.
• Growth — $499/mo — 10,000 interactions, custom voice profiles available.
• Scale — $1,499/mo — 50,000 interactions, custom voice cloning, dedicated voice tuning support.

14-day free trial, no credit card. 3-5 day setup.

Start your 14-day free trial →

Frequently asked questions

What is text to speech with emotion in 2026?

Text to speech with emotion in 2026 is generative audio that adjusts prosody, pacing, pitch, and warmth based on a directive — either a natural-language instruction ("speak warmly and slowly") or a structured emotion tag ("calm", "enthusiastic", "empathetic"). The output is rendered end-to-end by a generative model like GPT-Realtime-2, not assembled from phoneme samples. The result sounds like a person with a mood, not a flat voice with pitch contour applied.

What is the best text to speech model for a voice agent?

For voice agents in production, the best text to speech model in 2026 is GPT-Realtime-2 from OpenAI, with ElevenLabs v3 and Cartesia Sonic-2 as strong alternatives. GPT-Realtime-2 is the default at CallSphere because it integrates STT, reasoning, and TTS in a single model, which keeps turn latency under 800ms. ElevenLabs is the best for ultra-realistic single-voice work; Cartesia is the best for low-latency streaming use cases.

Can I use woman text to speech voices with emotion in CallSphere?

Yes. CallSphere ships 30+ female voices across 57+ languages with full emotion steering — warm, calm, confident, enthusiastic, professional. You pick the base voice in the admin console and add emotion directives in the agent prompt. Custom voice cloning is available on the Scale tier ($1,499/mo) with a 30-second sample and contractual voice licensing.

Is there a good adam text to speech voice for sales calls?

Yes. "Adam" is a common male voice profile name across ElevenLabs and other TTS engines. CallSphere ships multiple male voice options including a confident American male profile that we use for our default sales outbound agent. You can also clone a custom Adam-style voice on the Scale tier from a 30-second sample.

How does text to speech with characters work in 2026?

Text to speech with characters — meaning multiple distinct voice personas in one conversation or piece of content — is straightforward in 2026 generative models. You specify the voice profile per turn or per character tag. CallSphere supports this for hold messages, multi-language IVR replacement, and any use case where the agent needs to play multiple roles in one call. Most use cases are well served by a single voice with emotion steering across turns rather than multiple character voices.

What is the best speech to text model for paralinguistic features?

For paralinguistic features — pace, volume, hesitation, emotion in the caller's speech — the leading 2026 speech to text models are GPT-Realtime-Whisper, Deepgram Nova-3, and AssemblyAI Universal-2. CallSphere uses GPT-Realtime-Whisper because it integrates natively with GPT-Realtime-2 and shares the 128K context window, which lets the agent reason about caller emotion across the full conversation.

Can I get australian text to speech with emotion in CallSphere?

Yes. Australian English is one of the 57+ language and accent options. The same emotion steering directives work — warm, calm, confident, enthusiastic — and the model renders them in the Australian accent. We have customers using Australian voices for sales outbound to Australian markets and for hospitality concierge in Australian hotels.

How much does emotion-aware TTS cost in 2026?

GPT-Realtime-2's audio pricing is $32 per 1M input tokens, $64 per 1M output tokens, and $0.40 per 1M cached input tokens. A 10-minute voice call typically costs $0.30-$0.60 in raw model spend. CallSphere bundles this into our flat-rate plans ($149-$1,499/mo), so you do not see token-level pricing — you see interaction counts. A "Growth" customer at $499/mo gets 10,000 interactions including all emotion-aware TTS at no upcharge.

Related reading

• Siri voice generator — full 2026 guide
• WillFromAfar text to speech and other character voices
• Best text to speech app for engineers in 2026
• GPT-Realtime-2 128K context for voice agents
• Voice control android — 2026 capabilities